fs/erofs/zdata.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             https://www.huawei.com/
   5  */
   6 #include "zdata.h"
   7 #include "compress.h"
   8 #include <linux/prefetch.h>
   9
  10 #include <trace/events/erofs.h>
  11
  12 /*
  13  * since pclustersize is variable for big pcluster feature, introduce slab
  14  * pools implementation for different pcluster sizes.
  15  */
  16 struct z_erofs_pcluster_slab {
  17         struct kmem_cache *slab;
  18         unsigned int maxpages;
  19         char name[48];
  20 };
  21
  22 #define _PCLP(n) { .maxpages = n }
  23
  24 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
  25         _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
  26         _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
  27 };
  28
  29 static void z_erofs_destroy_pcluster_pool(void)
  30 {
  31         int i;
  32
  33         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  34                 if (!pcluster_pool[i].slab)
  35                         continue;
  36                 kmem_cache_destroy(pcluster_pool[i].slab);
  37                 pcluster_pool[i].slab = NULL;
  38         }
  39 }
  40
  41 static int z_erofs_create_pcluster_pool(void)
  42 {
  43         struct z_erofs_pcluster_slab *pcs;
  44         struct z_erofs_pcluster *a;
  45         unsigned int size;
  46
  47         for (pcs = pcluster_pool;
  48              pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
  49                 size = struct_size(a, compressed_pages, pcs->maxpages);
  50
  51                 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
  52                 pcs->slab = kmem_cache_create(pcs->name, size, 0,
  53                                               SLAB_RECLAIM_ACCOUNT, NULL);
  54                 if (pcs->slab)
  55                         continue;
  56
  57                 z_erofs_destroy_pcluster_pool();
  58                 return -ENOMEM;
  59         }
  60         return 0;
  61 }
  62
  63 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
  64 {
  65         int i;
  66
  67         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  68                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
  69                 struct z_erofs_pcluster *pcl;
  70
  71                 if (nrpages > pcs->maxpages)
  72                         continue;
  73
  74                 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
  75                 if (!pcl)
  76                         return ERR_PTR(-ENOMEM);
  77                 pcl->pclusterpages = nrpages;
  78                 return pcl;
  79         }
  80         return ERR_PTR(-EINVAL);
  81 }
  82
  83 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
  84 {
  85         int i;
  86
  87         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  88                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
  89
  90                 if (pcl->pclusterpages > pcs->maxpages)
  91                         continue;
  92
  93                 kmem_cache_free(pcs->slab, pcl);
  94                 return;
  95         }
  96         DBG_BUGON(1);
  97 }
  98
  99 /*
 100  * a compressed_pages[] placeholder in order to avoid
 101  * being filled with file pages for in-place decompression.
 102  */
 103 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
 104
 105 /* how to allocate cached pages for a pcluster */
 106 enum z_erofs_cache_alloctype {
 107         DONTALLOC,      /* don't allocate any cached pages */
 108         DELAYEDALLOC,   /* delayed allocation (at the time of submitting io) */
 109         /*
 110          * try to use cached I/O if page allocation succeeds or fallback
 111          * to in-place I/O instead to avoid any direct reclaim.
 112          */
 113         TRYALLOC,
 114 };
 115
 116 /*
 117  * tagged pointer with 1-bit tag for all compressed pages
 118  * tag 0 - the page is just found with an extra page reference
 119  */
 120 typedef tagptr1_t compressed_page_t;
 121
 122 #define tag_compressed_page_justfound(page) \
 123         tagptr_fold(compressed_page_t, page, 1)
 124
 125 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 126
 127 void z_erofs_exit_zip_subsystem(void)
 128 {
 129         destroy_workqueue(z_erofs_workqueue);
 130         z_erofs_destroy_pcluster_pool();
 131 }
 132
 133 static inline int z_erofs_init_workqueue(void)
 134 {
 135         const unsigned int onlinecpus = num_possible_cpus();
 136
 137         /*
 138          * no need to spawn too many threads, limiting threads could minimum
 139          * scheduling overhead, perhaps per-CPU threads should be better?
 140          */
 141         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
 142                                             WQ_UNBOUND | WQ_HIGHPRI,
 143                                             onlinecpus + onlinecpus / 4);
 144         return z_erofs_workqueue ? 0 : -ENOMEM;
 145 }
 146
 147 int __init z_erofs_init_zip_subsystem(void)
 148 {
 149         int err = z_erofs_create_pcluster_pool();
 150
 151         if (err)
 152                 return err;
 153         err = z_erofs_init_workqueue();
 154         if (err)
 155                 z_erofs_destroy_pcluster_pool();
 156         return err;
 157 }
 158
 159 enum z_erofs_collectmode {
 160         COLLECT_SECONDARY,
 161         COLLECT_PRIMARY,
 162         /*
 163          * The current collection was the tail of an exist chain, in addition
 164          * that the previous processed chained collections are all decided to
 165          * be hooked up to it.
 166          * A new chain will be created for the remaining collections which are
 167          * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
 168          * the next collection cannot reuse the whole page safely in
 169          * the following scenario:
 170          *  ________________________________________________________________
 171          * |      tail (partial) page     |       head (partial) page       |
 172          * |   (belongs to the next cl)   |   (belongs to the current cl)   |
 173          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 174          */
 175         COLLECT_PRIMARY_HOOKED,
 176         /*
 177          * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
 178          * could be dispatched into bypass queue later due to uptodated managed
 179          * pages. All related online pages cannot be reused for inplace I/O (or
 180          * pagevec) since it can be directly decoded without I/O submission.
 181          */
 182         COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
 183         /*
 184          * The current collection has been linked with the owned chain, and
 185          * could also be linked with the remaining collections, which means
 186          * if the processing page is the tail page of the collection, thus
 187          * the current collection can safely use the whole page (since
 188          * the previous collection is under control) for in-place I/O, as
 189          * illustrated below:
 190          *  ________________________________________________________________
 191          * |  tail (partial) page |          head (partial) page           |
 192          * |  (of the current cl) |      (of the previous collection)      |
 193          * |  PRIMARY_FOLLOWED or |                                        |
 194          * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
 195          *
 196          * [  (*) the above page can be used as inplace I/O.               ]
 197          */
 198         COLLECT_PRIMARY_FOLLOWED,
 199 };
 200
 201 struct z_erofs_collector {
 202         struct z_erofs_pagevec_ctor vector;
 203
 204         struct z_erofs_pcluster *pcl, *tailpcl;
 205         struct z_erofs_collection *cl;
 206         /* a pointer used to pick up inplace I/O pages */
 207         struct page **icpage_ptr;
 208         z_erofs_next_pcluster_t owned_head;
 209
 210         enum z_erofs_collectmode mode;
 211 };
 212
 213 struct z_erofs_decompress_frontend {
 214         struct inode *const inode;
 215
 216         struct z_erofs_collector clt;
 217         struct erofs_map_blocks map;
 218
 219         bool readahead;
 220         /* used for applying cache strategy on the fly */
 221         bool backmost;
 222         erofs_off_t headoffset;
 223 };
 224
 225 #define COLLECTOR_INIT() { \
 226         .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 227         .mode = COLLECT_PRIMARY_FOLLOWED }
 228
 229 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 230         .inode = __i, .clt = COLLECTOR_INIT(), \
 231         .backmost = true, }
 232
 233 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
 234 static DEFINE_MUTEX(z_pagemap_global_lock);
 235
 236 static void preload_compressed_pages(struct z_erofs_collector *clt,
 237                                      struct address_space *mc,
 238                                      enum z_erofs_cache_alloctype type,
 239                                      struct page **pagepool)
 240 {
 241         struct z_erofs_pcluster *pcl = clt->pcl;
 242         bool standalone = true;
 243         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 244                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 245         struct page **pages;
 246         pgoff_t index;
 247
 248         if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
 249                 return;
 250
 251         pages = pcl->compressed_pages;
 252         index = pcl->obj.index;
 253         for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
 254                 struct page *page;
 255                 compressed_page_t t;
 256                 struct page *newpage = NULL;
 257
 258                 /* the compressed page was loaded before */
 259                 if (READ_ONCE(*pages))
 260                         continue;
 261
 262                 page = find_get_page(mc, index);
 263
 264                 if (page) {
 265                         t = tag_compressed_page_justfound(page);
 266                 } else {
 267                         /* I/O is needed, no possible to decompress directly */
 268                         standalone = false;
 269                         switch (type) {
 270                         case DELAYEDALLOC:
 271                                 t = tagptr_init(compressed_page_t,
 272                                                 PAGE_UNALLOCATED);
 273                                 break;
 274                         case TRYALLOC:
 275                                 newpage = erofs_allocpage(pagepool, gfp);
 276                                 if (!newpage)
 277                                         continue;
 278                                 set_page_private(newpage,
 279                                                  Z_EROFS_PREALLOCATED_PAGE);
 280                                 t = tag_compressed_page_justfound(newpage);
 281                                 break;
 282                         default:        /* DONTALLOC */
 283                                 continue;
 284                         }
 285                 }
 286
 287                 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
 288                         continue;
 289
 290                 if (page)
 291                         put_page(page);
 292                 else if (newpage)
 293                         erofs_pagepool_add(pagepool, newpage);
 294         }
 295
 296         /*
 297          * don't do inplace I/O if all compressed pages are available in
 298          * managed cache since it can be moved to the bypass queue instead.
 299          */
 300         if (standalone)
 301                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 302 }
 303
 304 /* called by erofs_shrinker to get rid of all compressed_pages */
 305 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 306                                        struct erofs_workgroup *grp)
 307 {
 308         struct z_erofs_pcluster *const pcl =
 309                 container_of(grp, struct z_erofs_pcluster, obj);
 310         int i;
 311
 312         /*
 313          * refcount of workgroup is now freezed as 1,
 314          * therefore no need to worry about available decompression users.
 315          */
 316         for (i = 0; i < pcl->pclusterpages; ++i) {
 317                 struct page *page = pcl->compressed_pages[i];
 318
 319                 if (!page)
 320                         continue;
 321
 322                 /* block other users from reclaiming or migrating the page */
 323                 if (!trylock_page(page))
 324                         return -EBUSY;
 325
 326                 if (!erofs_page_is_managed(sbi, page))
 327                         continue;
 328
 329                 /* barrier is implied in the following 'unlock_page' */
 330                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 331                 detach_page_private(page);
 332                 unlock_page(page);
 333         }
 334         return 0;
 335 }
 336
 337 int erofs_try_to_free_cached_page(struct page *page)
 338 {
 339         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 340         int ret = 0;    /* 0 - busy */
 341
 342         if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
 343                 unsigned int i;
 344
 345                 for (i = 0; i < pcl->pclusterpages; ++i) {
 346                         if (pcl->compressed_pages[i] == page) {
 347                                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 348                                 ret = 1;
 349                                 break;
 350                         }
 351                 }
 352                 erofs_workgroup_unfreeze(&pcl->obj, 1);
 353
 354                 if (ret)
 355                         detach_page_private(page);
 356         }
 357         return ret;
 358 }
 359
 360 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 361 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
 362                                    struct page *page)
 363 {
 364         struct z_erofs_pcluster *const pcl = clt->pcl;
 365
 366         while (clt->icpage_ptr > pcl->compressed_pages)
 367                 if (!cmpxchg(--clt->icpage_ptr, NULL, page))
 368                         return true;
 369         return false;
 370 }
 371
 372 /* callers must be with collection lock held */
 373 static int z_erofs_attach_page(struct z_erofs_collector *clt,
 374                                struct page *page,
 375                                enum z_erofs_page_type type)
 376 {
 377         int ret;
 378
 379         /* give priority for inplaceio */
 380         if (clt->mode >= COLLECT_PRIMARY &&
 381             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 382             z_erofs_try_inplace_io(clt, page))
 383                 return 0;
 384
 385         ret = z_erofs_pagevec_enqueue(&clt->vector, page, type);
 386         clt->cl->vcnt += (unsigned int)ret;
 387
 388         return ret ? 0 : -EAGAIN;
 389 }
 390
 391 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt)
 392 {
 393         struct z_erofs_pcluster *pcl = clt->pcl;
 394         z_erofs_next_pcluster_t *owned_head = &clt->owned_head;
 395
 396         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 397         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 398                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 399                 *owned_head = &pcl->next;
 400                 /* so we can attach this pcluster to our submission chain. */
 401                 clt->mode = COLLECT_PRIMARY_FOLLOWED;
 402                 return;
 403         }
 404
 405         /*
 406          * type 2, link to the end of an existing open chain, be careful
 407          * that its submission is controlled by the original attached chain.
 408          */
 409         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 410                     *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
 411                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 412                 clt->mode = COLLECT_PRIMARY_HOOKED;
 413                 clt->tailpcl = NULL;
 414                 return;
 415         }
 416         /* type 3, it belongs to a chain, but it isn't the end of the chain */
 417         clt->mode = COLLECT_PRIMARY;
 418 }
 419
 420 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
 421                                      struct inode *inode,
 422                                      struct erofs_map_blocks *map)
 423 {
 424         struct z_erofs_pcluster *pcl = clt->pcl;
 425         struct z_erofs_collection *cl;
 426         unsigned int length;
 427
 428         /* to avoid unexpected loop formed by corrupted images */
 429         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 430                 DBG_BUGON(1);
 431                 return -EFSCORRUPTED;
 432         }
 433
 434         cl = z_erofs_primarycollection(pcl);
 435         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 436                 DBG_BUGON(1);
 437                 return -EFSCORRUPTED;
 438         }
 439
 440         length = READ_ONCE(pcl->length);
 441         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 442                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 443                         DBG_BUGON(1);
 444                         return -EFSCORRUPTED;
 445                 }
 446         } else {
 447                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 448
 449                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 450                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 451
 452                 while (llen > length &&
 453                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 454                         cpu_relax();
 455                         length = READ_ONCE(pcl->length);
 456                 }
 457         }
 458         mutex_lock(&cl->lock);
 459         /* used to check tail merging loop due to corrupted images */
 460         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 461                 clt->tailpcl = pcl;
 462
 463         z_erofs_try_to_claim_pcluster(clt);
 464         clt->cl = cl;
 465         return 0;
 466 }
 467
 468 static int z_erofs_register_collection(struct z_erofs_collector *clt,
 469                                        struct inode *inode,
 470                                        struct erofs_map_blocks *map)
 471 {
 472         struct z_erofs_pcluster *pcl;
 473         struct z_erofs_collection *cl;
 474         struct erofs_workgroup *grp;
 475         int err;
 476
 477         if (!(map->m_flags & EROFS_MAP_ENCODED)) {
 478                 DBG_BUGON(1);
 479                 return -EFSCORRUPTED;
 480         }
 481
 482         /* no available pcluster, let's allocate one */
 483         pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT);
 484         if (IS_ERR(pcl))
 485                 return PTR_ERR(pcl);
 486
 487         atomic_set(&pcl->obj.refcount, 1);
 488         pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 489         pcl->algorithmformat = map->m_algorithmformat;
 490         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 491                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 492                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 493
 494         /* new pclusters should be claimed as type 1, primary and followed */
 495         pcl->next = clt->owned_head;
 496         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 497
 498         cl = z_erofs_primarycollection(pcl);
 499         cl->pageofs = map->m_la & ~PAGE_MASK;
 500
 501         /*
 502          * lock all primary followed works before visible to others
 503          * and mutex_trylock *never* fails for a new pcluster.
 504          */
 505         mutex_init(&cl->lock);
 506         DBG_BUGON(!mutex_trylock(&cl->lock));
 507
 508         grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
 509         if (IS_ERR(grp)) {
 510                 err = PTR_ERR(grp);
 511                 goto err_out;
 512         }
 513
 514         if (grp != &pcl->obj) {
 515                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 516                 err = -EEXIST;
 517                 goto err_out;
 518         }
 519         /* used to check tail merging loop due to corrupted images */
 520         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 521                 clt->tailpcl = pcl;
 522         clt->owned_head = &pcl->next;
 523         clt->pcl = pcl;
 524         clt->cl = cl;
 525         return 0;
 526
 527 err_out:
 528         mutex_unlock(&cl->lock);
 529         z_erofs_free_pcluster(pcl);
 530         return err;
 531 }
 532
 533 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
 534                                    struct inode *inode,
 535                                    struct erofs_map_blocks *map)
 536 {
 537         struct erofs_workgroup *grp;
 538         int ret;
 539
 540         DBG_BUGON(clt->cl);
 541
 542         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
 543         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
 544         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 545
 546         if (!PAGE_ALIGNED(map->m_pa)) {
 547                 DBG_BUGON(1);
 548                 return -EINVAL;
 549         }
 550
 551         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
 552         if (grp) {
 553                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 554         } else {
 555                 ret = z_erofs_register_collection(clt, inode, map);
 556
 557                 if (!ret)
 558                         goto out;
 559                 if (ret != -EEXIST)
 560                         return ret;
 561         }
 562
 563         ret = z_erofs_lookup_collection(clt, inode, map);
 564         if (ret) {
 565                 erofs_workgroup_put(&clt->pcl->obj);
 566                 return ret;
 567         }
 568
 569 out:
 570         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 571                                   clt->cl->pagevec, clt->cl->vcnt);
 572
 573         /* since file-backed online pages are traversed in reverse order */
 574         clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages;
 575         return 0;
 576 }
 577
 578 /*
 579  * keep in mind that no referenced pclusters will be freed
 580  * only after a RCU grace period.
 581  */
 582 static void z_erofs_rcu_callback(struct rcu_head *head)
 583 {
 584         struct z_erofs_collection *const cl =
 585                 container_of(head, struct z_erofs_collection, rcu);
 586
 587         z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
 588                                            primary_collection));
 589 }
 590
 591 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 592 {
 593         struct z_erofs_pcluster *const pcl =
 594                 container_of(grp, struct z_erofs_pcluster, obj);
 595         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 596
 597         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 598 }
 599
 600 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 601 {
 602         struct z_erofs_pcluster *const pcl =
 603                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 604
 605         erofs_workgroup_put(&pcl->obj);
 606 }
 607
 608 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 609 {
 610         struct z_erofs_collection *cl = clt->cl;
 611
 612         if (!cl)
 613                 return false;
 614
 615         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 616         mutex_unlock(&cl->lock);
 617
 618         /*
 619          * if all pending pages are added, don't hold its reference
 620          * any longer if the pcluster isn't hosted by ourselves.
 621          */
 622         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 623                 z_erofs_collection_put(cl);
 624
 625         clt->cl = NULL;
 626         return true;
 627 }
 628
 629 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 630                                        unsigned int cachestrategy,
 631                                        erofs_off_t la)
 632 {
 633         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 634                 return false;
 635
 636         if (fe->backmost)
 637                 return true;
 638
 639         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 640                 la < fe->headoffset;
 641 }
 642
 643 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 644                                 struct page *page, struct page **pagepool)
 645 {
 646         struct inode *const inode = fe->inode;
 647         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 648         struct erofs_map_blocks *const map = &fe->map;
 649         struct z_erofs_collector *const clt = &fe->clt;
 650         const loff_t offset = page_offset(page);
 651         bool tight = true;
 652
 653         enum z_erofs_cache_alloctype cache_strategy;
 654         enum z_erofs_page_type page_type;
 655         unsigned int cur, end, spiltted, index;
 656         int err = 0;
 657
 658         /* register locked file pages as online pages in pack */
 659         z_erofs_onlinepage_init(page);
 660
 661         spiltted = 0;
 662         end = PAGE_SIZE;
 663 repeat:
 664         cur = end - 1;
 665
 666         /* lucky, within the range of the current map_blocks */
 667         if (offset + cur >= map->m_la &&
 668             offset + cur < map->m_la + map->m_llen) {
 669                 /* didn't get a valid collection previously (very rare) */
 670                 if (!clt->cl)
 671                         goto restart_now;
 672                 goto hitted;
 673         }
 674
 675         /* go ahead the next map_blocks */
 676         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 677
 678         if (z_erofs_collector_end(clt))
 679                 fe->backmost = false;
 680
 681         map->m_la = offset + cur;
 682         map->m_llen = 0;
 683         err = z_erofs_map_blocks_iter(inode, map, 0);
 684         if (err)
 685                 goto err_out;
 686
 687 restart_now:
 688         if (!(map->m_flags & EROFS_MAP_MAPPED))
 689                 goto hitted;
 690
 691         err = z_erofs_collector_begin(clt, inode, map);
 692         if (err)
 693                 goto err_out;
 694
 695         /* preload all compressed pages (maybe downgrade role if necessary) */
 696         if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy, map->m_la))
 697                 cache_strategy = TRYALLOC;
 698         else
 699                 cache_strategy = DONTALLOC;
 700
 701         preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 702                                  cache_strategy, pagepool);
 703
 704 hitted:
 705         /*
 706          * Ensure the current partial page belongs to this submit chain rather
 707          * than other concurrent submit chains or the noio(bypass) chain since
 708          * those chains are handled asynchronously thus the page cannot be used
 709          * for inplace I/O or pagevec (should be processed in strict order.)
 710          */
 711         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 712                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 713
 714         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 715         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 716                 zero_user_segment(page, cur, end);
 717                 goto next_part;
 718         }
 719
 720         /* let's derive page type */
 721         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 722                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 723                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 724                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 725
 726         if (cur)
 727                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 728
 729 retry:
 730         err = z_erofs_attach_page(clt, page, page_type);
 731         /* should allocate an additional short-lived page for pagevec */
 732         if (err == -EAGAIN) {
 733                 struct page *const newpage =
 734                                 alloc_page(GFP_NOFS | __GFP_NOFAIL);
 735
 736                 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
 737                 err = z_erofs_attach_page(clt, newpage,
 738                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 739                 if (!err)
 740                         goto retry;
 741         }
 742
 743         if (err)
 744                 goto err_out;
 745
 746         index = page->index - (map->m_la >> PAGE_SHIFT);
 747
 748         z_erofs_onlinepage_fixup(page, index, true);
 749
 750         /* bump up the number of spiltted parts of a page */
 751         ++spiltted;
 752         /* also update nr_pages */
 753         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 754 next_part:
 755         /* can be used for verification */
 756         map->m_llen = offset + cur - map->m_la;
 757
 758         end = cur;
 759         if (end > 0)
 760                 goto repeat;
 761
 762 out:
 763         z_erofs_onlinepage_endio(page);
 764
 765         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 766                   __func__, page, spiltted, map->m_llen);
 767         return err;
 768
 769         /* if some error occurred while processing this page */
 770 err_out:
 771         SetPageError(page);
 772         goto out;
 773 }
 774
 775 static void z_erofs_decompressqueue_work(struct work_struct *work);
 776 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 777                                        bool sync, int bios)
 778 {
 779         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
 780
 781         /* wake up the caller thread for sync decompression */
 782         if (sync) {
 783                 unsigned long flags;
 784
 785                 spin_lock_irqsave(&io->u.wait.lock, flags);
 786                 if (!atomic_add_return(bios, &io->pending_bios))
 787                         wake_up_locked(&io->u.wait);
 788                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 789                 return;
 790         }
 791
 792         if (atomic_add_return(bios, &io->pending_bios))
 793                 return;
 794         /* Use workqueue and sync decompression for atomic contexts only */
 795         if (in_atomic() || irqs_disabled()) {
 796                 queue_work(z_erofs_workqueue, &io->u.work);
 797                 sbi->opt.readahead_sync_decompress = true;
 798                 return;
 799         }
 800         z_erofs_decompressqueue_work(&io->u.work);
 801 }
 802
 803 static bool z_erofs_page_is_invalidated(struct page *page)
 804 {
 805         return !page->mapping && !z_erofs_is_shortlived_page(page);
 806 }
 807
 808 static void z_erofs_decompressqueue_endio(struct bio *bio)
 809 {
 810         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
 811         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
 812         blk_status_t err = bio->bi_status;
 813         struct bio_vec *bvec;
 814         struct bvec_iter_all iter_all;
 815
 816         bio_for_each_segment_all(bvec, bio, iter_all) {
 817                 struct page *page = bvec->bv_page;
 818
 819                 DBG_BUGON(PageUptodate(page));
 820                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 821
 822                 if (err)
 823                         SetPageError(page);
 824
 825                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
 826                         if (!err)
 827                                 SetPageUptodate(page);
 828                         unlock_page(page);
 829                 }
 830         }
 831         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
 832         bio_put(bio);
 833 }
 834
 835 static int z_erofs_decompress_pcluster(struct super_block *sb,
 836                                        struct z_erofs_pcluster *pcl,
 837                                        struct page **pagepool)
 838 {
 839         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 840         struct z_erofs_pagevec_ctor ctor;
 841         unsigned int i, inputsize, outputsize, llen, nr_pages;
 842         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 843         struct page **pages, **compressed_pages, *page;
 844
 845         enum z_erofs_page_type page_type;
 846         bool overlapped, partial;
 847         struct z_erofs_collection *cl;
 848         int err;
 849
 850         might_sleep();
 851         cl = z_erofs_primarycollection(pcl);
 852         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 853
 854         mutex_lock(&cl->lock);
 855         nr_pages = cl->nr_pages;
 856
 857         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 858                 pages = pages_onstack;
 859         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 860                    mutex_trylock(&z_pagemap_global_lock)) {
 861                 pages = z_pagemap_global;
 862         } else {
 863                 gfp_t gfp_flags = GFP_KERNEL;
 864
 865                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 866                         gfp_flags |= __GFP_NOFAIL;
 867
 868                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 869                                        gfp_flags);
 870
 871                 /* fallback to global pagemap for the lowmem scenario */
 872                 if (!pages) {
 873                         mutex_lock(&z_pagemap_global_lock);
 874                         pages = z_pagemap_global;
 875                 }
 876         }
 877
 878         for (i = 0; i < nr_pages; ++i)
 879                 pages[i] = NULL;
 880
 881         err = 0;
 882         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 883                                   cl->pagevec, 0);
 884
 885         for (i = 0; i < cl->vcnt; ++i) {
 886                 unsigned int pagenr;
 887
 888                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 889
 890                 /* all pages in pagevec ought to be valid */
 891                 DBG_BUGON(!page);
 892                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 893
 894                 if (z_erofs_put_shortlivedpage(pagepool, page))
 895                         continue;
 896
 897                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 898                         pagenr = 0;
 899                 else
 900                         pagenr = z_erofs_onlinepage_index(page);
 901
 902                 DBG_BUGON(pagenr >= nr_pages);
 903
 904                 /*
 905                  * currently EROFS doesn't support multiref(dedup),
 906                  * so here erroring out one multiref page.
 907                  */
 908                 if (pages[pagenr]) {
 909                         DBG_BUGON(1);
 910                         SetPageError(pages[pagenr]);
 911                         z_erofs_onlinepage_endio(pages[pagenr]);
 912                         err = -EFSCORRUPTED;
 913                 }
 914                 pages[pagenr] = page;
 915         }
 916         z_erofs_pagevec_ctor_exit(&ctor, true);
 917
 918         overlapped = false;
 919         compressed_pages = pcl->compressed_pages;
 920
 921         for (i = 0; i < pcl->pclusterpages; ++i) {
 922                 unsigned int pagenr;
 923
 924                 page = compressed_pages[i];
 925
 926                 /* all compressed pages ought to be valid */
 927                 DBG_BUGON(!page);
 928                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 929
 930                 if (!z_erofs_is_shortlived_page(page)) {
 931                         if (erofs_page_is_managed(sbi, page)) {
 932                                 if (!PageUptodate(page))
 933                                         err = -EIO;
 934                                 continue;
 935                         }
 936
 937                         /*
 938                          * only if non-head page can be selected
 939                          * for inplace decompression
 940                          */
 941                         pagenr = z_erofs_onlinepage_index(page);
 942
 943                         DBG_BUGON(pagenr >= nr_pages);
 944                         if (pages[pagenr]) {
 945                                 DBG_BUGON(1);
 946                                 SetPageError(pages[pagenr]);
 947                                 z_erofs_onlinepage_endio(pages[pagenr]);
 948                                 err = -EFSCORRUPTED;
 949                         }
 950                         pages[pagenr] = page;
 951
 952                         overlapped = true;
 953                 }
 954
 955                 /* PG_error needs checking for all non-managed pages */
 956                 if (PageError(page)) {
 957                         DBG_BUGON(PageUptodate(page));
 958                         err = -EIO;
 959                 }
 960         }
 961
 962         if (err)
 963                 goto out;
 964
 965         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 966         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 967                 outputsize = llen;
 968                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 969         } else {
 970                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 971                 partial = true;
 972         }
 973
 974         inputsize = pcl->pclusterpages * PAGE_SIZE;
 975         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 976                                         .sb = sb,
 977                                         .in = compressed_pages,
 978                                         .out = pages,
 979                                         .pageofs_out = cl->pageofs,
 980                                         .inputsize = inputsize,
 981                                         .outputsize = outputsize,
 982                                         .alg = pcl->algorithmformat,
 983                                         .inplace_io = overlapped,
 984                                         .partial_decoding = partial
 985                                  }, pagepool);
 986
 987 out:
 988         /* must handle all compressed pages before ending pages */
 989         for (i = 0; i < pcl->pclusterpages; ++i) {
 990                 page = compressed_pages[i];
 991
 992                 if (erofs_page_is_managed(sbi, page))
 993                         continue;
 994
 995                 /* recycle all individual short-lived pages */
 996                 (void)z_erofs_put_shortlivedpage(pagepool, page);
 997
 998                 WRITE_ONCE(compressed_pages[i], NULL);
 999         }
1000
1001         for (i = 0; i < nr_pages; ++i) {
1002                 page = pages[i];
1003                 if (!page)
1004                         continue;
1005
1006                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1007
1008                 /* recycle all individual short-lived pages */
1009                 if (z_erofs_put_shortlivedpage(pagepool, page))
1010                         continue;
1011
1012                 if (err < 0)
1013                         SetPageError(page);
1014
1015                 z_erofs_onlinepage_endio(page);
1016         }
1017
1018         if (pages == z_pagemap_global)
1019                 mutex_unlock(&z_pagemap_global_lock);
1020         else if (pages != pages_onstack)
1021                 kvfree(pages);
1022
1023         cl->nr_pages = 0;
1024         cl->vcnt = 0;
1025
1026         /* all cl locks MUST be taken before the following line */
1027         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1028
1029         /* all cl locks SHOULD be released right now */
1030         mutex_unlock(&cl->lock);
1031
1032         z_erofs_collection_put(cl);
1033         return err;
1034 }
1035
1036 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1037                                      struct page **pagepool)
1038 {
1039         z_erofs_next_pcluster_t owned = io->head;
1040
1041         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1042                 struct z_erofs_pcluster *pcl;
1043
1044                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1045                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1046
1047                 /* no possible that 'owned' equals NULL */
1048                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1049
1050                 pcl = container_of(owned, struct z_erofs_pcluster, next);
1051                 owned = READ_ONCE(pcl->next);
1052
1053                 z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1054         }
1055 }
1056
1057 static void z_erofs_decompressqueue_work(struct work_struct *work)
1058 {
1059         struct z_erofs_decompressqueue *bgq =
1060                 container_of(work, struct z_erofs_decompressqueue, u.work);
1061         struct page *pagepool = NULL;
1062
1063         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1064         z_erofs_decompress_queue(bgq, &pagepool);
1065
1066         erofs_release_pages(&pagepool);
1067         kvfree(bgq);
1068 }
1069
1070 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1071                                                unsigned int nr,
1072                                                struct page **pagepool,
1073                                                struct address_space *mc,
1074                                                gfp_t gfp)
1075 {
1076         const pgoff_t index = pcl->obj.index;
1077         bool tocache = false;
1078
1079         struct address_space *mapping;
1080         struct page *oldpage, *page;
1081
1082         compressed_page_t t;
1083         int justfound;
1084
1085 repeat:
1086         page = READ_ONCE(pcl->compressed_pages[nr]);
1087         oldpage = page;
1088
1089         if (!page)
1090                 goto out_allocpage;
1091
1092         /*
1093          * the cached page has not been allocated and
1094          * an placeholder is out there, prepare it now.
1095          */
1096         if (page == PAGE_UNALLOCATED) {
1097                 tocache = true;
1098                 goto out_allocpage;
1099         }
1100
1101         /* process the target tagged pointer */
1102         t = tagptr_init(compressed_page_t, page);
1103         justfound = tagptr_unfold_tags(t);
1104         page = tagptr_unfold_ptr(t);
1105
1106         /*
1107          * preallocated cached pages, which is used to avoid direct reclaim
1108          * otherwise, it will go inplace I/O path instead.
1109          */
1110         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1111                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1112                 set_page_private(page, 0);
1113                 tocache = true;
1114                 goto out_tocache;
1115         }
1116         mapping = READ_ONCE(page->mapping);
1117
1118         /*
1119          * file-backed online pages in plcuster are all locked steady,
1120          * therefore it is impossible for `mapping' to be NULL.
1121          */
1122         if (mapping && mapping != mc)
1123                 /* ought to be unmanaged pages */
1124                 goto out;
1125
1126         /* directly return for shortlived page as well */
1127         if (z_erofs_is_shortlived_page(page))
1128                 goto out;
1129
1130         lock_page(page);
1131
1132         /* only true if page reclaim goes wrong, should never happen */
1133         DBG_BUGON(justfound && PagePrivate(page));
1134
1135         /* the page is still in manage cache */
1136         if (page->mapping == mc) {
1137                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1138
1139                 ClearPageError(page);
1140                 if (!PagePrivate(page)) {
1141                         /*
1142                          * impossible to be !PagePrivate(page) for
1143                          * the current restriction as well if
1144                          * the page is already in compressed_pages[].
1145                          */
1146                         DBG_BUGON(!justfound);
1147
1148                         justfound = 0;
1149                         set_page_private(page, (unsigned long)pcl);
1150                         SetPagePrivate(page);
1151                 }
1152
1153                 /* no need to submit io if it is already up-to-date */
1154                 if (PageUptodate(page)) {
1155                         unlock_page(page);
1156                         page = NULL;
1157                 }
1158                 goto out;
1159         }
1160
1161         /*
1162          * the managed page has been truncated, it's unsafe to
1163          * reuse this one, let's allocate a new cache-managed page.
1164          */
1165         DBG_BUGON(page->mapping);
1166         DBG_BUGON(!justfound);
1167
1168         tocache = true;
1169         unlock_page(page);
1170         put_page(page);
1171 out_allocpage:
1172         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1173         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1174                 erofs_pagepool_add(pagepool, page);
1175                 cond_resched();
1176                 goto repeat;
1177         }
1178 out_tocache:
1179         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1180                 /* turn into temporary page if fails (1 ref) */
1181                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1182                 goto out;
1183         }
1184         attach_page_private(page, pcl);
1185         /* drop a refcount added by allocpage (then we have 2 refs here) */
1186         put_page(page);
1187
1188 out:    /* the only exit (for tracing and debugging) */
1189         return page;
1190 }
1191
1192 static struct z_erofs_decompressqueue *
1193 jobqueue_init(struct super_block *sb,
1194               struct z_erofs_decompressqueue *fgq, bool *fg)
1195 {
1196         struct z_erofs_decompressqueue *q;
1197
1198         if (fg && !*fg) {
1199                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1200                 if (!q) {
1201                         *fg = true;
1202                         goto fg_out;
1203                 }
1204                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1205         } else {
1206 fg_out:
1207                 q = fgq;
1208                 init_waitqueue_head(&fgq->u.wait);
1209                 atomic_set(&fgq->pending_bios, 0);
1210         }
1211         q->sb = sb;
1212         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1213         return q;
1214 }
1215
1216 /* define decompression jobqueue types */
1217 enum {
1218         JQ_BYPASS,
1219         JQ_SUBMIT,
1220         NR_JOBQUEUES,
1221 };
1222
1223 static void *jobqueueset_init(struct super_block *sb,
1224                               struct z_erofs_decompressqueue *q[],
1225                               struct z_erofs_decompressqueue *fgq, bool *fg)
1226 {
1227         /*
1228          * if managed cache is enabled, bypass jobqueue is needed,
1229          * no need to read from device for all pclusters in this queue.
1230          */
1231         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1232         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1233
1234         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1235 }
1236
1237 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1238                                     z_erofs_next_pcluster_t qtail[],
1239                                     z_erofs_next_pcluster_t owned_head)
1240 {
1241         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1242         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1243
1244         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1245         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1246                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1247
1248         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1249
1250         WRITE_ONCE(*submit_qtail, owned_head);
1251         WRITE_ONCE(*bypass_qtail, &pcl->next);
1252
1253         qtail[JQ_BYPASS] = &pcl->next;
1254 }
1255
1256 static void z_erofs_submit_queue(struct super_block *sb,
1257                                  struct z_erofs_decompress_frontend *f,
1258                                  struct page **pagepool,
1259                                  struct z_erofs_decompressqueue *fgq,
1260                                  bool *force_fg)
1261 {
1262         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1263         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1264         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1265         void *bi_private;
1266         z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1267         /* bio is NULL initially, so no need to initialize last_{index,bdev} */
1268         pgoff_t last_index;
1269         struct block_device *last_bdev;
1270         unsigned int nr_bios = 0;
1271         struct bio *bio = NULL;
1272
1273         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1274         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1275         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1276
1277         /* by default, all need io submission */
1278         q[JQ_SUBMIT]->head = owned_head;
1279
1280         do {
1281                 struct erofs_map_dev mdev;
1282                 struct z_erofs_pcluster *pcl;
1283                 pgoff_t cur, end;
1284                 unsigned int i = 0;
1285                 bool bypass = true;
1286
1287                 /* no possible 'owned_head' equals the following */
1288                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1289                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1290
1291                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1292
1293                 /* no device id here, thus it will always succeed */
1294                 mdev = (struct erofs_map_dev) {
1295                         .m_pa = blknr_to_addr(pcl->obj.index),
1296                 };
1297                 (void)erofs_map_dev(sb, &mdev);
1298
1299                 cur = erofs_blknr(mdev.m_pa);
1300                 end = cur + pcl->pclusterpages;
1301
1302                 /* close the main owned chain at first */
1303                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1304                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1305
1306                 do {
1307                         struct page *page;
1308
1309                         page = pickup_page_for_submission(pcl, i++, pagepool,
1310                                                           MNGD_MAPPING(sbi),
1311                                                           GFP_NOFS);
1312                         if (!page)
1313                                 continue;
1314
1315                         if (bio && (cur != last_index + 1 ||
1316                                     last_bdev != mdev.m_bdev)) {
1317 submit_bio_retry:
1318                                 submit_bio(bio);
1319                                 bio = NULL;
1320                         }
1321
1322                         if (!bio) {
1323                                 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS);
1324                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1325
1326                                 bio_set_dev(bio, mdev.m_bdev);
1327                                 last_bdev = mdev.m_bdev;
1328                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1329                                         LOG_SECTORS_PER_BLOCK;
1330                                 bio->bi_private = bi_private;
1331                                 bio->bi_opf = REQ_OP_READ;
1332                                 if (f->readahead)
1333                                         bio->bi_opf |= REQ_RAHEAD;
1334                                 ++nr_bios;
1335                         }
1336
1337                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1338                                 goto submit_bio_retry;
1339
1340                         last_index = cur;
1341                         bypass = false;
1342                 } while (++cur < end);
1343
1344                 if (!bypass)
1345                         qtail[JQ_SUBMIT] = &pcl->next;
1346                 else
1347                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1348         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1349
1350         if (bio)
1351                 submit_bio(bio);
1352
1353         /*
1354          * although background is preferred, no one is pending for submission.
1355          * don't issue workqueue for decompression but drop it directly instead.
1356          */
1357         if (!*force_fg && !nr_bios) {
1358                 kvfree(q[JQ_SUBMIT]);
1359                 return;
1360         }
1361         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1362 }
1363
1364 static void z_erofs_runqueue(struct super_block *sb,
1365                              struct z_erofs_decompress_frontend *f,
1366                              struct page **pagepool, bool force_fg)
1367 {
1368         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1369
1370         if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1371                 return;
1372         z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1373
1374         /* handle bypass queue (no i/o pclusters) immediately */
1375         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1376
1377         if (!force_fg)
1378                 return;
1379
1380         /* wait until all bios are completed */
1381         io_wait_event(io[JQ_SUBMIT].u.wait,
1382                       !atomic_read(&io[JQ_SUBMIT].pending_bios));
1383
1384         /* handle synchronous decompress queue in the caller context */
1385         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1386 }
1387
1388 /*
1389  * Since partial uptodate is still unimplemented for now, we have to use
1390  * approximate readmore strategies as a start.
1391  */
1392 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1393                                       struct readahead_control *rac,
1394                                       erofs_off_t end,
1395                                       struct page **pagepool,
1396                                       bool backmost)
1397 {
1398         struct inode *inode = f->inode;
1399         struct erofs_map_blocks *map = &f->map;
1400         erofs_off_t cur;
1401         int err;
1402
1403         if (backmost) {
1404                 map->m_la = end;
1405                 err = z_erofs_map_blocks_iter(inode, map,
1406                                               EROFS_GET_BLOCKS_READMORE);
1407                 if (err)
1408                         return;
1409
1410                 /* expend ra for the trailing edge if readahead */
1411                 if (rac) {
1412                         loff_t newstart = readahead_pos(rac);
1413
1414                         cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1415                         readahead_expand(rac, newstart, cur - newstart);
1416                         return;
1417                 }
1418                 end = round_up(end, PAGE_SIZE);
1419         } else {
1420                 end = round_up(map->m_la, PAGE_SIZE);
1421
1422                 if (!map->m_llen)
1423                         return;
1424         }
1425
1426         cur = map->m_la + map->m_llen - 1;
1427         while (cur >= end) {
1428                 pgoff_t index = cur >> PAGE_SHIFT;
1429                 struct page *page;
1430
1431                 page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1432                 if (!page)
1433                         goto skip;
1434
1435                 if (PageUptodate(page)) {
1436                         unlock_page(page);
1437                         put_page(page);
1438                         goto skip;
1439                 }
1440
1441                 err = z_erofs_do_read_page(f, page, pagepool);
1442                 if (err)
1443                         erofs_err(inode->i_sb,
1444                                   "readmore error at page %lu @ nid %llu",
1445                                   index, EROFS_I(inode)->nid);
1446                 put_page(page);
1447 skip:
1448                 if (cur < PAGE_SIZE)
1449                         break;
1450                 cur = (index << PAGE_SHIFT) - 1;
1451         }
1452 }
1453
1454 static int z_erofs_readpage(struct file *file, struct page *page)
1455 {
1456         struct inode *const inode = page->mapping->host;
1457         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1458         struct page *pagepool = NULL;
1459         int err;
1460
1461         trace_erofs_readpage(page, false);
1462         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1463
1464         z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1465                                   &pagepool, true);
1466         err = z_erofs_do_read_page(&f, page, &pagepool);
1467         z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1468
1469         (void)z_erofs_collector_end(&f.clt);
1470
1471         /* if some compressed cluster ready, need submit them anyway */
1472         z_erofs_runqueue(inode->i_sb, &f, &pagepool, true);
1473
1474         if (err)
1475                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1476
1477         if (f.map.mpage)
1478                 put_page(f.map.mpage);
1479
1480         erofs_release_pages(&pagepool);
1481         return err;
1482 }
1483
1484 static void z_erofs_readahead(struct readahead_control *rac)
1485 {
1486         struct inode *const inode = rac->mapping->host;
1487         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1488         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1489         struct page *pagepool = NULL, *head = NULL, *page;
1490         unsigned int nr_pages;
1491
1492         f.readahead = true;
1493         f.headoffset = readahead_pos(rac);
1494
1495         z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1496                                   readahead_length(rac) - 1, &pagepool, true);
1497         nr_pages = readahead_count(rac);
1498         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1499
1500         while ((page = readahead_page(rac))) {
1501                 set_page_private(page, (unsigned long)head);
1502                 head = page;
1503         }
1504
1505         while (head) {
1506                 struct page *page = head;
1507                 int err;
1508
1509                 /* traversal in reverse order */
1510                 head = (void *)page_private(page);
1511
1512                 err = z_erofs_do_read_page(&f, page, &pagepool);
1513                 if (err)
1514                         erofs_err(inode->i_sb,
1515                                   "readahead error at page %lu @ nid %llu",
1516                                   page->index, EROFS_I(inode)->nid);
1517                 put_page(page);
1518         }
1519         z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1520         (void)z_erofs_collector_end(&f.clt);
1521
1522         z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1523                          sbi->opt.readahead_sync_decompress &&
1524                          nr_pages <= sbi->opt.max_sync_decompress_pages);
1525         if (f.map.mpage)
1526                 put_page(f.map.mpage);
1527         erofs_release_pages(&pagepool);
1528 }
1529
1530 const struct address_space_operations z_erofs_aops = {
1531         .readpage = z_erofs_readpage,
1532         .readahead = z_erofs_readahead,
1533 };